Advancing terrestrial laser scanning for 3D classification of Spanish moss (Tillandsia usneoides L.) using unsupervised, machine learning, and deep learning methods.

Quantifying the distribution of Spanish moss (Tillandsia usneoides L.) is challenging because it grows suspended from high tree branches, limiting manual sampling. Terrestrial laser scanning (TLS) provides a non-destructive means of capturing vegetation structure in three dimensions. However, no established methods exist for identifying Spanish moss from TLS data. We evaluated five classification methods for distinguishing Spanish moss in TLS-derived point cloud data: Graph, DBSCAN, Random Forest (RF), Kernel Point Convolution (KPConv), and PointNet++. PointNet++ achieved the highest accuracy (81%), followed by DBSCAN (70%), KPConv (61%), RF (54%), and Graph (52%). Unsupervised methods required minimal computational resources (2-3 min, 8-16 GB memory) without training. RF required 3 h for training, 8 for prediction with 1024 GB memory. Deep learning methods required substantially more: KPConv needed 60 h for training, 4 for prediction (256 GB), while PointNet++ required 48 h for training, 1 for prediction (128 GB). Agreement was lowest in the central and upper canopy due to occlusion. Surface variation, PCA1, and verticality contributed most to accurate predictions. These results demonstrate the feasibility of using TLS and advanced classification methods for non-destructive Spanish moss mapping and highlight the accurate classification ability of PointNet++ for future biomass estimation at landscape scales.